@bunshin-cc/bunshin 0.1.4 → 0.2.0

package/README.md

@@ -1,6 +1,6 @@
 # Bunshin
 
-Bunshin lets your company’s AI agents learn from each other. Agents save useful facts and workflows from their work, then recall relevant lessons from the company workspace.
+Bunshin lets your company’s AI agents (and robots) learn from each other. Your repo’s own SPEC.md, TESTING.md and SKILL.md sections, other teams’ rules and what earlier sessions learned reach the agent when they matter; what a session learns is tried locally, then reviewed and shared.
 
 ## Connect a project
 
@@ -12,13 +12,13 @@ cd /path/to/your/project
 bunshin setup --harness codex
 ```
 
-In Bunshin, open **Agent keys** and create a key for this computer. Paste it into the hidden setup prompt. Setup reads your role, team, and permissions from the workspace, saves the key in a private file outside the project, and configures four project hooks. It does not need a separate model API key.
+In Bunshin, open **Agent keys** and create a key for this computer. Paste it into the hidden setup prompt. Setup reads your role, team, and permissions from the workspace, saves the key in a private file outside the project, and configures the project hooks (five for Codex, six plus an MCP server for Claude Code). It does not need a separate model API key.
 
-Run `codex` interactively in the same project. Review and accept its project trust prompt, then run `/hooks` and review and trust the four Bunshin hooks. Project trust and hook trust are separate steps. If `/hooks` does not list Bunshin, check that you opened the connected project and trusted its configuration. This guide tests interactive sessions; `codex exec` is not a substitute for this setup check.
+Run `codex` interactively in the same project. Review and accept its project trust prompt, then run `/hooks` and review and trust the five Bunshin hooks. Project trust and hook trust are separate steps. If `/hooks` does not list Bunshin, check that you opened the connected project and trusted its configuration. This guide tests interactive sessions; `codex exec` is not a substitute for this setup check.
 
 For a linked Git worktree, setup puts Codex hooks in the primary checkout, where Codex reads them. The printed path identifies that shared hook file. Each connected worktree keeps its own `.bunshin` store and private connection. Worktrees created from a bare repository are not supported yet; use a full checkout.
 
-For Claude Code, use `bunshin setup --harness claude-code`, then start a new session in the project.
+For Claude Code, use `bunshin setup --harness claude-code`, then start a new session in the project and approve the `bunshin` MCP server it adds to `.mcp.json` (it gives the model the `bunshin_recall` and `bunshin_load` tools). `/learn` in the Claude Code plugin, or `bunshin learn "<what to remember>"`, captures knowledge on demand.
 
 For Pi, use Node.js 22.19 or later and run `bunshin setup --harness pi`. It installs a project extension in `.pi/extensions/bunshin.js`. Start Pi in the project (or run `/reload` in an existing session), review its project resources, and check `/bunshin status`. The extension ships in this npm package; you do not need a second package or a repository checkout.
 
@@ -27,10 +27,11 @@ For Pi, use Node.js 22.19 or later and run `bunshin setup --harness pi`. It inst
 ```sh
 bunshin status
 bunshin sync --json
-bunshin recall "What should I know for my next task?"
+bunshin index
+bunshin recall "changing a state transition other teams depend on" --hypothetical "the owning team must approve first"
 ```
 
-An empty workspace has no lessons to recall yet. Work with your agent, teach it a reusable project fact or procedure, and end the session. Reflection runs in the background. Check `bunshin lessons list` and your workspace’s **Knowledge** page, then ask a relevant question in a fresh agent session.
+`bunshin index` shows how the repo’s docs split into sections and which paths push each one. Learned knowledge arrives after real sessions: when a session was worth it, a reflection runs at its end (for Claude Code, as a fork of the session, so it reuses the prompt cache). Suggestions that were not applied automatically wait in `bunshin review`.
 
 `bunshin demo` runs an offline example using temporary synthetic data. It does not verify your workspace connection or automatic learning.